Research map / Label-efficient and robust learning / Recognition and long-tail
Temperature in Cosine-based Softmax Loss
Takumi Kobayashi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use least-squares kappa instead of rule-based kappa | cosine softmax ResNet-50 · ImageNet | accuracy: 3.36 -> 77.27 | Improved |
| use least-squares kappa instead of maximum-likelihood kappa | cosine softmax ResNet-50 · ImageNet | accuracy: 0.10 -> 77.27 | Improved |
| use LS kappa instead of trainable kappa | cosine softmax ResNet-50 · ImageNet | accuracy: 75.51 -> 77.27 | Improved |
| set kappa to 2kappa* instead of softmax | ResNet-50 · ImageNet | accuracy: 77.12 -> 77.81 | Improved |
| raise fixed kappa from 5 to 20 (ablation) | cosine softmax ResNet-50 · ImageNet | accuracy: 68.95 -> 77.76 | Improved |
| use kappa=60 instead of kappa* (ablation) | cosine softmax · ImageNet | accuracy: 77.27 -> 76.88 | Hurt |
| larger kappa=60 vs softmax (ablation) | ResNet-50 pretrained · CUB-200 linear probe | accuracy: 69.05 -> 75.63 | Improved |
| larger kappa=60 vs softmax (ablation) | ResNet-50 pretrained · Flower-102 linear probe | accuracy: 86.39 -> 92.19 | Improved |
| larger kappa=60 vs softmax (ablation) | ResNet-50 pretrained · CUB-200 fine-tuning | accuracy: 80.17 -> 81.18 | Improved |
| small kappa=10 vs softmax (ablation) | ResNet-50 pretrained · CUB-200 linear probe | accuracy: 69.05 -> 47.13 | Hurt |
| use LS kappa in long-tailed training | ResNet-50 · iNaturalist2018 | accuracy: 66.76 -> 67.75 | Improved |
| use fixed kappa=10 instead of LS kappa*=18.54 (ablation) | ResNet-50 · iNaturalist2018 | accuracy: 67.75 -> 57.36 | Hurt |
Closest papers by meaning
- Large-Margin Softmax Loss for Convolutional Neural Networks (ICML 2016 · similarity 0.85)
- On Temperature Scaling and Conformal Prediction of Deep Classifiers (ICML 2025 · similarity 0.85)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (NeurIPS 2024 · similarity 0.83)
- To Cool or not to Cool? Temperature Network Meets Large Foundation Models via DRO (ICML 2024 · similarity 0.83)
- On the Limitations of Temperature Scaling for Distributions with Overlaps (ICLR 2024 · similarity 0.82)
- Building Vision Models upon Heat Conduction (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
